Skip to main content

cmd/datanode/
objbench.rs

1// Copyright 2023 Greptime Team
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7//     http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15use std::path::PathBuf;
16use std::sync::Arc;
17use std::time::Instant;
18
19use clap::Parser;
20use colored::Colorize;
21use futures::stream;
22use mito2::access_layer::{
23    AccessLayer, AccessLayerRef, Metrics, OperationType, SstWriteRequest, WriteType,
24};
25use mito2::cache::{CacheManager, CacheManagerRef};
26use mito2::config::{FulltextIndexConfig, MitoConfig, Mode};
27use mito2::read::FlatSource;
28use mito2::sst::FormatType;
29use mito2::sst::file::{FileHandle, FileMeta};
30use mito2::sst::file_purger::{FilePurger, FilePurgerRef};
31use mito2::sst::index::intermediate::IntermediateManager;
32use mito2::sst::index::puffin_manager::PuffinManagerFactory;
33use mito2::sst::parquet::WriteOptions;
34use mito2::sst::parquet::reader::ParquetReaderBuilder;
35use mito2::worker::write_cache_from_config;
36use object_store::ObjectStore;
37use parquet::file::metadata::FooterTail;
38use regex::Regex;
39use snafu::OptionExt;
40use store_api::path_utils::region_name;
41use store_api::region_request::PathType;
42use store_api::storage::FileId;
43
44use crate::datanode::tool_util::{
45    build_object_store, extract_region_metadata, max_row_group_uncompressed_size, parse_config,
46};
47use crate::error;
48
49/// Object storage benchmark command
50#[derive(Debug, Parser)]
51pub struct ObjbenchCommand {
52    /// Path to the object-store config file (TOML). Must deserialize into object_store::config::ObjectStoreConfig.
53    #[clap(long, value_name = "FILE")]
54    pub config: PathBuf,
55
56    /// Source SST file path in object-store (e.g. "region_dir/<uuid>.parquet").
57    #[clap(long, value_name = "PATH")]
58    pub source: String,
59
60    /// Verbose output
61    #[clap(short, long, default_value_t = false)]
62    pub verbose: bool,
63
64    /// Output file path for pprof flamegraph (enables profiling)
65    #[clap(long, value_name = "FILE")]
66    pub pprof_file: Option<PathBuf>,
67}
68
69impl ObjbenchCommand {
70    pub async fn run(&self) -> error::Result<()> {
71        if self.verbose {
72            common_telemetry::init_default_ut_logging();
73        }
74
75        println!("{}", "Starting objbench with config:".cyan().bold());
76
77        // Build object store from config
78        let (store_cfg, mut mito_engine_config, _wal_config) = parse_config(&self.config)?;
79
80        let object_store = build_object_store(&store_cfg).await?;
81        println!("{} Object store initialized", "✓".green());
82
83        // Prepare source identifiers
84        let components = parse_file_dir_components(&self.source)?;
85        println!(
86            "{} Source path parsed: {}, components: {:?}",
87            "✓".green(),
88            self.source,
89            components
90        );
91
92        // Load parquet metadata to extract RegionMetadata and file stats
93        println!("{}", "Loading parquet metadata...".yellow());
94        let file_size = object_store
95            .stat(&self.source)
96            .await
97            .map_err(|e| {
98                error::IllegalConfigSnafu {
99                    msg: format!("stat failed: {e}"),
100                }
101                .build()
102            })?
103            .content_length();
104        let parquet_meta = load_parquet_metadata(object_store.clone(), &self.source, file_size)
105            .await
106            .map_err(|e| {
107                error::IllegalConfigSnafu {
108                    msg: format!("read parquet metadata failed: {e}"),
109                }
110                .build()
111            })?;
112
113        let region_meta = extract_region_metadata(&self.source, &parquet_meta)?;
114        let num_rows = parquet_meta.file_metadata().num_rows() as u64;
115        let num_row_groups = parquet_meta.num_row_groups() as u64;
116        let max_row_group_uncompressed_size = max_row_group_uncompressed_size(&parquet_meta);
117
118        println!(
119            "{} Metadata loaded - rows: {}, size: {} bytes",
120            "✓".green(),
121            num_rows,
122            file_size
123        );
124
125        // Build a FileHandle for the source file
126        let file_meta = FileMeta {
127            region_id: region_meta.region_id,
128            file_id: components.file_id,
129            time_range: Default::default(),
130            level: 0,
131            file_size,
132            max_row_group_uncompressed_size,
133            available_indexes: Default::default(),
134            indexes: Default::default(),
135            index_file_size: 0,
136            index_version: 0,
137            num_rows,
138            num_row_groups,
139            sequence: None,
140            partition_expr: None,
141            num_series: 0,
142            ..Default::default()
143        };
144        let src_handle = FileHandle::new(file_meta, new_noop_file_purger());
145
146        // Build the reader for a single file via ParquetReaderBuilder
147        let table_dir = components.table_dir();
148        let (src_access_layer, cache_manager) = build_access_layer_simple(
149            &components,
150            object_store.clone(),
151            &mut mito_engine_config,
152            &store_cfg.data_home,
153        )
154        .await?;
155        let reader_build_start = Instant::now();
156
157        let reader = ParquetReaderBuilder::new(
158            table_dir,
159            components.path_type,
160            src_handle.clone(),
161            object_store.clone(),
162        )
163        .expected_metadata(Some(region_meta.clone()))
164        .build()
165        .await
166        .map_err(|e| {
167            error::IllegalConfigSnafu {
168                msg: format!("build reader failed: {e:?}"),
169            }
170            .build()
171        })?;
172        let reader = reader.ok_or_else(|| {
173            error::IllegalConfigSnafu {
174                msg: format!(
175                    "build reader returned no readable rows for source file {}",
176                    src_handle.file_id()
177                ),
178            }
179            .build()
180        })?;
181
182        let reader_build_elapsed = reader_build_start.elapsed();
183        let total_rows = reader.parquet_metadata().file_metadata().num_rows();
184        println!("{} Reader built in {:?}", "✓".green(), reader_build_elapsed);
185        let reader_stream = Box::pin(stream::try_unfold(reader, |mut reader| async move {
186            let batch = reader.next_record_batch().await?;
187            Ok(batch.map(|batch| (batch, reader)))
188        }));
189
190        // Build write request
191        let fulltext_index_config = FulltextIndexConfig {
192            create_on_compaction: Mode::Disable,
193            ..Default::default()
194        };
195
196        let source =
197            FlatSource::new_stream(region_meta.schema.arrow_schema().clone(), reader_stream);
198        let write_req = SstWriteRequest {
199            op_type: OperationType::Flush,
200            metadata: region_meta,
201            source,
202            cache_manager,
203            storage: None,
204            max_sequence: None,
205            sst_write_format: FormatType::PrimaryKey,
206            index_options: Default::default(),
207            index_config: mito_engine_config.index.clone(),
208            inverted_index_config: MitoConfig::default().inverted_index,
209            fulltext_index_config,
210            bloom_filter_index_config: MitoConfig::default().bloom_filter_index,
211            preserve_row_sequence: false,
212        };
213
214        // Write SST
215        println!("{}", "Writing SST...".yellow());
216
217        // Start profiling if pprof_file is specified
218        #[cfg(unix)]
219        let profiler_guard = if self.pprof_file.is_some() {
220            println!("{} Starting profiling...", "⚡".yellow());
221            Some(
222                pprof::ProfilerGuardBuilder::default()
223                    .frequency(99)
224                    .blocklist(&["libc", "libgcc", "pthread", "vdso"])
225                    .build()
226                    .map_err(|e| {
227                        error::IllegalConfigSnafu {
228                            msg: format!("Failed to start profiler: {e}"),
229                        }
230                        .build()
231                    })?,
232            )
233        } else {
234            None
235        };
236
237        #[cfg(not(unix))]
238        if self.pprof_file.is_some() {
239            eprintln!(
240                "{}: Profiling is not supported on this platform",
241                "Warning".yellow()
242            );
243        }
244
245        let write_start = Instant::now();
246        let mut metrics = Metrics::new(WriteType::Flush);
247        let infos = src_access_layer
248            .write_sst(write_req, &WriteOptions::default(), &mut metrics)
249            .await
250            .map_err(|e| {
251                error::IllegalConfigSnafu {
252                    msg: format!("write_sst failed: {e:?}"),
253                }
254                .build()
255            })?;
256
257        let write_elapsed = write_start.elapsed();
258
259        // Stop profiling and generate flamegraph if enabled
260        #[cfg(unix)]
261        if let (Some(guard), Some(pprof_file)) = (profiler_guard, &self.pprof_file) {
262            println!("{} Generating flamegraph...", "🔥".yellow());
263            match guard.report().build() {
264                Ok(report) => {
265                    let mut flamegraph_data = Vec::new();
266                    if let Err(e) = report.flamegraph(&mut flamegraph_data) {
267                        println!("{}: Failed to generate flamegraph: {}", "Error".red(), e);
268                    } else if let Err(e) = std::fs::write(pprof_file, flamegraph_data) {
269                        println!(
270                            "{}: Failed to write flamegraph to {}: {}",
271                            "Error".red(),
272                            pprof_file.display(),
273                            e
274                        );
275                    } else {
276                        println!(
277                            "{} Flamegraph saved to {}",
278                            "✓".green(),
279                            pprof_file.display().to_string().cyan()
280                        );
281                    }
282                }
283                Err(e) => {
284                    println!("{}: Failed to generate pprof report: {}", "Error".red(), e);
285                }
286            }
287        }
288        assert_eq!(infos.len(), 1);
289        let dst_file_id = infos[0].file_id;
290        let dst_file_path = format!("{}/{}.parquet", components.region_dir(), dst_file_id);
291        let mut dst_index_path = None;
292        if infos[0].index_metadata.file_size > 0 {
293            dst_index_path = Some(format!(
294                "{}/index/{}.puffin",
295                components.region_dir(),
296                dst_file_id
297            ));
298        }
299
300        // Report results with ANSI colors
301        println!("\n{} {}", "Write complete!".green().bold(), "✓".green());
302        println!("  {}: {}", "Destination file".bold(), dst_file_path.cyan());
303        println!("  {}: {}", "Rows".bold(), total_rows.to_string().cyan());
304        println!(
305            "  {}: {}",
306            "File size".bold(),
307            format!("{} bytes", file_size).cyan()
308        );
309        println!(
310            "  {}: {:?}",
311            "Reader build time".bold(),
312            reader_build_elapsed
313        );
314        println!("  {}: {:?}", "Total time".bold(), write_elapsed);
315
316        // Print metrics in a formatted way
317        println!("  {}: {:?}", "Metrics".bold(), metrics,);
318
319        // Print infos
320        println!("  {}: {:?}", "Index".bold(), infos[0].index_metadata);
321
322        // Cleanup
323        println!("\n{}", "Cleaning up...".yellow());
324        object_store.delete(&dst_file_path).await.map_err(|e| {
325            error::IllegalConfigSnafu {
326                msg: format!("Failed to delete dest file {}: {}", dst_file_path, e),
327            }
328            .build()
329        })?;
330        println!("{} Temporary file {} deleted", "✓".green(), dst_file_path);
331
332        if let Some(index_path) = dst_index_path {
333            object_store.delete(&index_path).await.map_err(|e| {
334                error::IllegalConfigSnafu {
335                    msg: format!("Failed to delete dest index file {}: {}", index_path, e),
336                }
337                .build()
338            })?;
339            println!(
340                "{} Temporary index file {} deleted",
341                "✓".green(),
342                index_path
343            );
344        }
345
346        println!("\n{}", "Benchmark completed successfully!".green().bold());
347        Ok(())
348    }
349}
350
351#[derive(Debug)]
352struct FileDirComponents {
353    catalog: String,
354    schema: String,
355    table_id: u32,
356    region_sequence: u32,
357    path_type: PathType,
358    file_id: FileId,
359}
360
361impl FileDirComponents {
362    fn table_dir(&self) -> String {
363        format!("data/{}/{}/{}", self.catalog, self.schema, self.table_id)
364    }
365
366    fn region_dir(&self) -> String {
367        let region_name = region_name(self.table_id, self.region_sequence);
368        match self.path_type {
369            PathType::Bare => {
370                format!(
371                    "data/{}/{}/{}/{}",
372                    self.catalog, self.schema, self.table_id, region_name
373                )
374            }
375            PathType::Data => {
376                format!(
377                    "data/{}/{}/{}/{}/data",
378                    self.catalog, self.schema, self.table_id, region_name
379                )
380            }
381            PathType::Metadata => {
382                format!(
383                    "data/{}/{}/{}/{}/metadata",
384                    self.catalog, self.schema, self.table_id, region_name
385                )
386            }
387        }
388    }
389}
390
391fn parse_file_dir_components(path: &str) -> error::Result<FileDirComponents> {
392    // Define the regex pattern to match all three path styles
393    let pattern =
394        r"^data/([^/]+)/([^/]+)/([^/]+)/([^/]+)_([^/]+)(?:/data|/metadata)?/(.+).parquet$";
395
396    // Compile the regex
397    let re = Regex::new(pattern).expect("Invalid regex pattern");
398
399    // Determine the path type
400    let path_type = if path.contains("/data/") {
401        PathType::Data
402    } else if path.contains("/metadata/") {
403        PathType::Metadata
404    } else {
405        PathType::Bare
406    };
407
408    // Try to match the path
409    let components = (|| {
410        let captures = re.captures(path)?;
411        if captures.len() != 7 {
412            return None;
413        }
414        let mut components = FileDirComponents {
415            catalog: "".to_string(),
416            schema: "".to_string(),
417            table_id: 0,
418            region_sequence: 0,
419            path_type,
420            file_id: FileId::default(),
421        };
422        // Extract the components
423        components.catalog = captures.get(1)?.as_str().to_string();
424        components.schema = captures.get(2)?.as_str().to_string();
425        components.table_id = captures[3].parse().ok()?;
426        components.region_sequence = captures[5].parse().ok()?;
427        let file_id_str = &captures[6];
428        components.file_id = FileId::parse_str(file_id_str).ok()?;
429        Some(components)
430    })();
431    components.context(error::IllegalConfigSnafu {
432        msg: format!("Expect valid source file path, got: {}", path),
433    })
434}
435
436async fn build_access_layer_simple(
437    components: &FileDirComponents,
438    object_store: ObjectStore,
439    config: &mut MitoConfig,
440    data_home: &str,
441) -> error::Result<(AccessLayerRef, CacheManagerRef)> {
442    let _ = config.index.sanitize(data_home, &config.inverted_index);
443    let puffin_manager = PuffinManagerFactory::new(
444        &config.index.aux_path,
445        config.index.staging_size.as_bytes(),
446        Some(config.index.write_buffer_size.as_bytes() as _),
447        config.index.staging_ttl,
448    )
449    .await
450    .map_err(|e| {
451        error::IllegalConfigSnafu {
452            msg: format!("Failed to build access layer: {e:?}"),
453        }
454        .build()
455    })?;
456
457    let intermediate_manager = IntermediateManager::init_fs(&config.index.aux_path)
458        .await
459        .map_err(|e| {
460            error::IllegalConfigSnafu {
461                msg: format!("Failed to build IntermediateManager: {e:?}"),
462            }
463            .build()
464        })?
465        .with_buffer_size(Some(config.index.write_buffer_size.as_bytes() as _));
466
467    let cache_manager =
468        build_cache_manager(config, puffin_manager.clone(), intermediate_manager.clone()).await?;
469    let layer = AccessLayer::new(
470        components.table_dir(),
471        components.path_type,
472        object_store,
473        puffin_manager,
474        intermediate_manager,
475    );
476    Ok((Arc::new(layer), cache_manager))
477}
478
479async fn build_cache_manager(
480    config: &MitoConfig,
481    puffin_manager: PuffinManagerFactory,
482    intermediate_manager: IntermediateManager,
483) -> error::Result<CacheManagerRef> {
484    let write_cache = write_cache_from_config(config, puffin_manager, intermediate_manager, None)
485        .await
486        .map_err(|e| {
487            error::IllegalConfigSnafu {
488                msg: format!("Failed to build write cache: {e:?}"),
489            }
490            .build()
491        })?;
492    let cache_manager = Arc::new(
493        CacheManager::builder()
494            .sst_meta_cache_size(config.sst_meta_cache_size.as_bytes())
495            .vector_cache_size(config.vector_cache_size.as_bytes())
496            .page_cache_size(config.page_cache_size.as_bytes())
497            .selector_result_cache_size(config.selector_result_cache_size.as_bytes())
498            .range_result_cache_size(config.range_result_cache_size.as_bytes())
499            .prefilter_result_cache_size(config.prefilter_result_cache_size.as_bytes())
500            .index_metadata_size(config.index.metadata_cache_size.as_bytes())
501            .index_content_size(config.index.content_cache_size.as_bytes())
502            .index_content_page_size(config.index.content_cache_page_size.as_bytes())
503            .index_result_cache_size(config.index.result_cache_size.as_bytes())
504            .puffin_metadata_size(config.index.metadata_cache_size.as_bytes())
505            .write_cache(write_cache)
506            .build(),
507    );
508    Ok(cache_manager)
509}
510
511fn new_noop_file_purger() -> FilePurgerRef {
512    #[derive(Debug)]
513    struct Noop;
514    impl FilePurger for Noop {
515        fn remove_file(&self, _file_meta: FileMeta, _is_delete: bool, _index_outdated: bool) {}
516    }
517    Arc::new(Noop)
518}
519
520async fn load_parquet_metadata(
521    object_store: ObjectStore,
522    path: &str,
523    file_size: u64,
524) -> Result<parquet::file::metadata::ParquetMetaData, Box<dyn std::error::Error + Send + Sync>> {
525    use parquet::file::FOOTER_SIZE;
526    use parquet::file::metadata::ParquetMetaDataReader;
527    let actual_size = if file_size == 0 {
528        object_store.stat(path).await?.content_length()
529    } else {
530        file_size
531    };
532    if actual_size < FOOTER_SIZE as u64 {
533        return Err("file too small".into());
534    }
535    let prefetch: u64 = 64 * 1024;
536    let start = actual_size.saturating_sub(prefetch);
537    let buffer = object_store
538        .read_with(path)
539        .range(start..actual_size)
540        .await?
541        .to_vec();
542    let buffer_len = buffer.len();
543    let mut footer = [0; 8];
544    footer.copy_from_slice(&buffer[buffer_len - FOOTER_SIZE..]);
545    let footer = FooterTail::try_new(&footer)?;
546    let metadata_len = footer.metadata_length() as u64;
547    if actual_size - (FOOTER_SIZE as u64) < metadata_len {
548        return Err("invalid footer/metadata length".into());
549    }
550    if (metadata_len as usize) <= buffer_len - FOOTER_SIZE {
551        let metadata_start = buffer_len - metadata_len as usize - FOOTER_SIZE;
552        let meta = ParquetMetaDataReader::decode_metadata(
553            &buffer[metadata_start..buffer_len - FOOTER_SIZE],
554        )?;
555        Ok(meta)
556    } else {
557        let metadata_start = actual_size - metadata_len - FOOTER_SIZE as u64;
558        let data = object_store
559            .read_with(path)
560            .range(metadata_start..(actual_size - FOOTER_SIZE as u64))
561            .await?
562            .to_vec();
563        let meta = ParquetMetaDataReader::decode_metadata(&data)?;
564        Ok(meta)
565    }
566}
567
568#[cfg(test)]
569mod tests {
570    use std::path::PathBuf;
571    use std::str::FromStr;
572
573    use common_base::readable_size::ReadableSize;
574    use store_api::region_request::PathType;
575
576    use crate::datanode::objbench::parse_file_dir_components;
577    use crate::datanode::tool_util::parse_config;
578
579    #[test]
580    fn test_parse_dir() {
581        let meta_path = "data/greptime/public/1024/1024_0000000000/metadata/00020380-009c-426d-953e-b4e34c15af34.parquet";
582        let c = parse_file_dir_components(meta_path).unwrap();
583        assert_eq!(
584            c.file_id.to_string(),
585            "00020380-009c-426d-953e-b4e34c15af34"
586        );
587        assert_eq!(c.catalog, "greptime");
588        assert_eq!(c.schema, "public");
589        assert_eq!(c.table_id, 1024);
590        assert_eq!(c.region_sequence, 0);
591        assert_eq!(c.path_type, PathType::Metadata);
592
593        let c = parse_file_dir_components(
594            "data/greptime/public/1024/1024_0000000000/data/00020380-009c-426d-953e-b4e34c15af34.parquet",
595        ).unwrap();
596        assert_eq!(
597            c.file_id.to_string(),
598            "00020380-009c-426d-953e-b4e34c15af34"
599        );
600        assert_eq!(c.catalog, "greptime");
601        assert_eq!(c.schema, "public");
602        assert_eq!(c.table_id, 1024);
603        assert_eq!(c.region_sequence, 0);
604        assert_eq!(c.path_type, PathType::Data);
605
606        let c = parse_file_dir_components(
607            "data/greptime/public/1024/1024_0000000000/00020380-009c-426d-953e-b4e34c15af34.parquet",
608        ).unwrap();
609        assert_eq!(
610            c.file_id.to_string(),
611            "00020380-009c-426d-953e-b4e34c15af34"
612        );
613        assert_eq!(c.catalog, "greptime");
614        assert_eq!(c.schema, "public");
615        assert_eq!(c.table_id, 1024);
616        assert_eq!(c.region_sequence, 0);
617        assert_eq!(c.path_type, PathType::Bare);
618    }
619
620    #[test]
621    fn test_parse_config() {
622        let path = "../../config/datanode.example.toml";
623        let (storage, engine, _wal) = parse_config(&PathBuf::from_str(path).unwrap()).unwrap();
624        assert_eq!(storage.data_home, "./greptimedb_data");
625        assert_eq!(engine.index.staging_size, ReadableSize::gb(2));
626    }
627}